Interaxon, the Canadian company behind the Muse sleep and meditation headband, told Science magazine it had used about 50,000 hours of EEG recordings, most collected from device users, to train an AI model. Its privacy policy had not explicitly told users about that use, the investigation found.
Key facts
- Interaxon trained a “brain foundation model” on about 50,000 hours of EEG recordings, most from users of its devices (Malwarebytes, reporting Science).
- After Science asked questions, Interaxon updated its privacy policy to say de-identified data is used to train machine-learning models.
- Nadia Kumentas, Interaxon’s chief marketing officer, acknowledged that users should be explicitly informed.
EEG, short for electroencephalography, records the brain’s electrical activity through sensors on the scalp. The Muse headband uses it to track brain activity during sleep and meditation.
A foundation model is a large, general-purpose AI model that can be adapted for many tasks. Training one takes a lot of data, which is why a library of real brain recordings is valuable.
Brain data can reveal more than most people expect. As consumer neurotech spreads, who may use that data, and for what, is a question for users and regulators everywhere, including in Nigeria, Kenya and South Africa.
What Science found
Interaxon told Science it used about 50,000 hours of EEG recordings to train the model, most of them collected from Muse users, Medical Daily reported. The figure comes from the company and hasn’t been independently audited. Users were not explicitly told about the training.
The company changed its privacy policy after Science’s questions. The policy now says de-identified data is used to train machine-learning models.
Interaxon removed names and email addresses from the recordings before training, Malwarebytes reported. But once data has shaped a trained model, removing its influence is very difficult.
What Interaxon says
Kumentas acknowledged that users should be explicitly informed, according to Malwarebytes. Muse’s current privacy policy lets users withdraw consent for new third-party research sharing, but does not state an opt-out for Interaxon’s own model training. It says account deletion removes personal information while de-identified and aggregated data may continue to be used; it does not say past model training can be undone.
Muse is not alone
Science reviewed 57 companies that collect brain data, Malwarebytes reported. Only six explicitly said they would not sell users’ brain data.
More than two-thirds allow brain data to pass to a new owner if the company is acquired. Eleven had a privacy policy that covered only website visitors, and four had no policy at all.
This fits an earlier finding. A 2024 Neurorights Foundation report found that 29 of 30 consumer neurotech companies it surveyed appeared to have access to consumers’ neural data, with “no meaningful limitations to this access.”
Why brainwave data is sensitive
Research suggests brain recordings can identify people and hold sensitive personal information, Malwarebytes noted:
- A 2018 Korea University study of 17 people identified them from resting-state EEG 88.4% of the time.
- A 2026 University of South-Eastern Norway study found EEG data can hold sensitive, patient-specific information, and that removing identifiers may not stop re-identification.
These studies do not show what can be inferred from Muse data specifically. Consumer headbands have fewer sensors than clinical systems, Malwarebytes cautioned.
What it means for users in Nigeria, Kenya and South Africa
In the US, Colorado and California now treat neural data as sensitive, according to Malwarebytes. We could not confirm whether any African data protection law names brainwave or neural data specifically.
Until the rules are clear, the burden sits with the buyer. Before using any brain-sensing wearable, Malwarebytes suggests asking the maker:
- Is my EEG data used to train AI?
- Can I opt out?
- Can I delete recordings you have already collected?
- What happens to my data if the company is sold or goes bust?
If the answers are no, or there is no answer at all, it may be worth reconsidering the device.
What’s next
Watch whether Interaxon offers a specific opt-out for its own model training and explains what, if anything, can be removed from earlier de-identified recordings or trained models, and whether data protection regulators, including Africa’s, start naming neural data in their rules.







